Layered Video Objects Detection Based on LBP and Codebook

Background subtraction is often one of the first tasks and a critical part of the machine vision systems. Background modeling methods that only utilize image blocks or pixels suffer from unacceptable false negative detecting rate. A novel layered background modeling method is proposed in objects detecting. First, every block on the first layer is modeled via texture based on local binary pattern (LBP) operators. Then the modeling granularity is deflated onto the second layer to model via codebook. Layered match is done from top down when a new video frame enters. Experimental results prove that the proposed novel approach is useful and robust. It efficiently avoids the false negative detection rate in the pixel-based background modeling when the object color is similar to the background, and also stops the false positives occurring on the contour areas of the moving objects due to the block model.